Junjie Hu

Ruixue Lian, William A. Sethares, Junjie Hu

Supervised contrastive learning (SCL) frameworks treat each class as independent and thus consider all classes to be equally important. This neglects the common scenario in which label hierarchy exists, where fine-grained classes under the same category show more similarity than very different ones. This paper introduces a family of Label-Aware SCL methods (LASCL) that incorporates hierarchical information to SCL by leveraging similarities between classes, resulting in creating a more well-structured and discriminative feature space. This is achieved by first adjusting the distance between instances based on measures of the proximity of their classes with the scaled instance-instance-wise contrastive. An additional instance-center-wise contrastive is introduced to move within-class examples closer to their centers, which are represented by a set of learnable label parameters. The learned label parameters can be directly used as a nearest neighbor classifier without further finetuning. In this way, a better feature representation is generated with improvements of intra-cluster compactness and inter-cluster separation. Experiments on three datasets show that the proposed LASCL works well on text classification of distinguishing a single label among multi-labels, outperforming the baseline supervised approaches. Our code is publicly available.

Cheng Luo, Zefan Cai, Hanshi Sun et al.

Transformer-based large language models (LLMs) demonstrate impressive performance in long context generation. Extending the context length has disproportionately shifted the memory footprint of LLMs during inference to the key-value cache (KV cache). In this paper, we propose HEADINFER, which offloads the KV cache to CPU RAM while avoiding the need to fully store the KV cache for any transformer layer on the GPU. HEADINFER employs a fine-grained, head-wise offloading strategy, maintaining only selective attention heads KV cache on the GPU while computing attention output dynamically. Through roofline analysis, we demonstrate that HEADINFER maintains computational efficiency while significantly reducing memory footprint. We evaluate HEADINFER on the Llama-3-8B model with a 1-million-token sequence, reducing the GPU memory footprint of the KV cache from 128 GB to 1 GB and the total GPU memory usage from 207 GB to 17 GB, achieving a 92% reduction compared to BF16 baseline inference. Notably, HEADINFER enables 4-million-token inference with an 8B model on a single consumer GPU with 24GB memory (e.g., NVIDIA RTX 4090) without approximation methods.